Date of Award

5-23-2010

Document Type

Honors Thesis

Department

Computer Science

First Advisor

John P. MacCormick

Language

English

Abstract

Image segmentation into superpixels is a common early step in computer vision algorithms. Several algorithms already exist that produce good quality results. PathFinder is a competing image segmentation algorithm that creates high quality superpixels at faster speeds. Our goal is to transfer computation from the CPU to the GPU, which has the potential to execute parallel tasks much more efficiently. We investigate CUDA, a freely available platform for programming GPUs, and learn what kinds of operations benefit most from this approach. We find a 3 to 5 times speed up of the PathFinder algorithm overall, with an improvement of three orders of magnitude in some components of the algorithm.

COinS